Before you buy AI marking: what the DfE's EdTech evidence pilot means for school leaders

As of 7 October 2026, the DfE-funded EdTech Evidence Board pilot, delivered by the Chartered College of Teaching, has given no product an award, because the evidence wasn't strong enough. Its report was published on 24 September 2026. For schools, that means the burden of proof sits with the buyer.

If you have sat through an AI marking demo this year, you will know the pitch: hours saved, marks that match an examiner's, teachers who love it. What you rarely see is the evidence behind those claims. The pilot doesn't settle whether EdTech works, but it does show school leaders what to ask for.

It is worth being precise, because this was not a classroom trial of AI marking. No schools tested products, and nobody measured marking accuracy or workload. Instead, suppliers submitted portfolios of evidence, and independent reviewers judged them against a published set of criteria, after a set of baseline checks covering areas such as data protection, safety and accessibility.

Phase two ran from April 2025 to April 2026. According to the report, "a total of 96 EOIs were received", and "the pilot proceeded with 12 products - four for the initial review pathway and eight for the full review pathway". Notably for anyone considering AI marking, the products placed on the initial review pathway "all utilised generative AI in some way", because the delivery team saw a risk in admitting AI products to the full review at this stage.

Don't confuse it with the DfE's EdTech Testbed (in-school evaluations) or the AI tools for education contracts (prototypes). None is a seal of approval.

The headline is in the outcome. The delivery team, its board and the DfE took "a joint decision ... not to confer formal awards at this stage", because "the overall distribution and consistency of evidence against the criteria were not yet sufficient to support fair, reliable and comparable awards across the cohort" (p.37).

The detail is more useful still. On the claims suppliers make publicly, reviewers found that "in all but a few cases it was impossible to substantiate these claims as the research or evidence was not shared" (p.16). Some suppliers "believed they met baseline requirements but were unable to demonstrate this" (p.6).

On the type of evidence, the report is blunt: "weaker performing evidence types included case studies, analytics and product feedback, surveys and correlational studies" (p.43). And on whether products had been tested across a range of users and contexts, "only one of the portfolios met this subcriterion" (p.42).

For AI specifically, the conclusion is clear. "The pace of product development and deployment in this area has, in most cases, outstripped the generation of robust evidence relating to impact" (p.58). For school leaders, the report's own advice is simple: "adopt a critical approach when evaluating product claims" (p.53).

The Chartered College has published an EdTech Check tool for educators, and it is worth using. Alongside it, these five questions will tell you a lot in a single conversation. If you are buying across a trust, our MAT guide to procuring AI marking tools covers the wider process.

1. What exactly is the claim, and how was it measured? "Saves teachers hours" is not a claim you can test. "Reduced marking time on a Year 11 maths mock by X minutes per script, measured across these schools" is. Ask for the study, not the slide. If they can't show you how it was measured, be cautious.

2. Accurate compared with what? Agreement with a single teacher, with a senior examiner, or with the final moderated mark are very different benchmarks. Ask how close the AI's marks were, on which questions and which papers, and where it disagreed. Extended responses and follow-through marks are where tools tend to struggle, so ask about those specifically.

3. Is there evidence from schools like yours? Given that only one portfolio met the range-of-contexts subcriterion, this is the question most likely to expose a gap. A tool tested in one trust, one subject or one cohort may behave differently with your students' handwriting, your exam boards and your mix of prior attainment.

4. Where does the teacher sit? Ask to see, step by step, how a teacher reviews, overrides and signs off a mark. Who owns the final grade? What happens when the AI is unsure? Our earlier piece on keeping humans in the loop goes into what good looks like here.

5. Can they show the basics? Data protection, safety, accessibility and security are the baseline the pilot checked first. Ask for the paperwork, check it against the DfE's generative AI product safety expectations, and ask which AI model sits underneath and which version your evidence relates to. If you are still setting out the rules for AI use in school, our guide to AI policy for schools covers the governance side.

The most persuasive evidence will be the evidence you gather yourself. If you are trialling AI marking for schools, run a small, honest pilot: for AI mock exam marking, mark a sample of mock scripts both ways, compare the marks, time the process, and ask your team what changed. At Excelas, our position is simple: AI supports the teacher's marking, it never replaces it, and a teacher reviews and owns every final mark.

A useful benchmark comes from JCQ. Its instructions for conducting non-examination assessment state that "teachers must not use AI as the sole or primary means of marking candidates' work". That rule applies to NEA rather than internal mocks, but it is a sensible standard to hold any tool to.

The pilot didn't tell us which products to trust. It told us how to decide. Ask for the evidence, check it against your own context, and keep teachers firmly in charge of the marks. That is a good test for any tool, including ours. Ask us the same five questions: our accuracy and timing figures are on our AI mock exam marking page, or bring one of your own scripts to a 20-minute call and we'll mark it live.